2x integer key decoding perf

Summary:
In my testing this reduces the overhead of integer decoding by 2x. The idea of the fix is straight-forward:
* Don't call Field* to get information, instead cache as much information as possible in our Rdb_field_packing, similar to InnoDB's row_prebuilt_t. This has a few advantages:
  * Avoid expensive virtual function calls
  * More cache friendly
* Similarly, don't use field->move_field / move_field_offset  to move field to the target buffer and then back - just write to the buffer using pre-calculated field offset directly
* Use template integer unpacking function to enable compiler to unroll expensive loops into a constant number of mov instructions.

NOTE: FDO data needs to be updated for this diff in order to take full advantage of this change.

Reviewed By: hermanlee

Differential Revision: D24383422

fbshipit-source-id: e00ee825935

storage/rocksdb/rdb_converter.cc

@@ -817,7 +817,7 @@ int Rdb_converter::convert_record_from_storage_format(
     return err;
   }
 
-  if (!decode_value) {
+  if (!decode_value || get_decode_fields()->size() == 0) {
     // We are done
     return HA_EXIT_SUCCESS;
   }